Abstract: Breakthrough research drives greatly advances in science and technology, and expands the boundaries of human knowledge. The early identification of such achievements is essential for forward-looking basic research planning, the efficient allocation of scientific resources, and the formulation of national innovation strategies. However, existing studies on the identification of breakthrough achievements often focus on a single-dimension rather than an integrated theoretical framework, and few adequately explore the early formation mechanisms of breakthrough achievements. To address these limitations, this study proposes a theoretical framework that grounded in knowledge flow theory, encompassing the process of knowledge input, knowledge production, and knowledge output. Based on this framework, a multi-dimensional early identification index system comprising 15 indicators is constructed by integrating features across the three stages. Multiple machine learning algorithms are then employed to build identification models, from which the most effective model is selected. Finally, SHAP is applied to interpret the model and to quantify the relative importance and contributions of different features in the identification process. The results indicate that: ① the CatBoost model demonstrates the best performance in early identification. ② three key early signals are particularly influential, which are citations within five years, disruption index within five years in the knowledge output stage, and the author's highest academic achievement in the knowledge production stage. ③ the model exhibits strong generalization capability when it verified by APS milestone papers. Overall, this study proposes a novel paradigm that integrates predictive accuracy with interpretability for the early identification of breakthrough research, and provides evidence for understanding its early formation mechanisms.

Key words: Breakthrough achievements, Knowledge flow theory, Early identification, Machine learning, SHAP